Vāda
Deliberation Teams
The room outperforms the individual.
Vāda is not a factory.Vāda is deliberation.
The AI industry is racing to build better factories. Agent swarms that browse the web, read codebases, execute tasks at lightspeed. These are execution tools, and they have a blind spot. They assume the question is the right question. Give an execution swarm a bad idea and it will execute it perfectly.
Vāda sits above execution. You bring a decision to Vāda before you commit resources to it — to stress-test the strategy, attack the assumptions, and find the blind spots.
No external tools. No web access. No code execution. The friction of the debate must not be diluted.
unexamined.
down; execution noise
cannot pass up.
the swarm.
- Structured recommendation
- Convergence and dissent, named
- What was assumed and what remains open
Deep thinking, on demand.
You bring a question. Vāda launches a team of agents on it — reviewers running across multiple models, working through the question under a structured protocol. No single model's blind spots. No one viewpoint deciding for you.
The shape of the work is defined by the team you launch. A team can be a panel of independent reviewers — each on a different model — returning their analysis. Or a deeper engagement with multiple rounds, cross-critique, an audit gate, and a revision loop. You pick the team that fits the decision.
Whatever the team's shape, the deliverable is the same: a structured conclusion with the full transcript attached. Every reviewer's reasoning, preserved. Auditable. Defensible to anyone who later asks how the decision was reached.
The question you walked in with is not always the question you walk out with. That clarity is what you came for.
Built on theAtta Engine.
The Atta Engine is the deliberation runtime built by AttaLabs. Vāda is the first system to run on it. The engine handles parallel execution, state passing, audit gates, revision loops, and the audit trail. Vāda provides the teams.
A team is defined in YAML. A team is a sequence of rounds. A round has agents. Agents have models, prompts, and a place in the flow. To launch a new team, write a YAML file. To change a team, change the YAML. There is no team-specific code.
The same engine runs every team — from a single reviewer on one model to a multi-round panel with cross-critique and audit. Whatever the YAML says, the engine runs.
The engine is the foundation. The teams are how you put it to work.
Choose your models.
Frontier, open, or mixed. Claude as your Critic, Llama as your Strategist. The room doesn’t care.
schema_version: "2.0"
id: example-team
name: An example team
agents:
- name: agent_a
model: anthropic/claude-sonnet
- name: agent_b
model: openai/gpt-4
- name: agent_c
model: google/gemini
flow:
rounds:
- id: open
layout: parallel
agents: [agent_a, agent_b, agent_c]
message_template: "{{question}}"
- id: respond
layout: parallel
agents: [agent_a, agent_b, agent_c]
message_template: "{{previousRound}}"
- id: synthesize
layout: sequential
agents: [agent_a]
message_template: "{{allPreviousOutputs}}"
- ■ Structured conclusion
- ■ Full transcript
- ■ Cost & timing metrics
Vāda is not a chatbot.
No memory. No personality. No ongoing relationship.
Vāda is not a swarm.
No tools. No file access. No code execution. Agents think. They don't act.
Vāda is not a workflow.
No steps. No automation. No integrations.
Vāda is not trying to be helpful.
It is trying to be right.
Vāda is one system from AttaLabs.More are on the way.
AttaLabs is the lab. Vāda is the first system to ship from it — built on the Atta Engine, the deliberation runtime the lab maintains. Two further systems are in design.
